multi_brains Framework Source Code Analysis

1. Course Content

• This section analyzes the core functionalities of the multi_brains functional package architecture.
• In subsequent lessons, each new action function introduced in the tutorials will be explained separately.

[!TIP]

• This course is a purely code-analysis course, intended for users who want a deeper understanding of the functionality. If you only need to experience the functionality, you can skip this section.

2. Source Code Package Structure

2.1 Functional Package File Structure

├── config
│   ├── map_mapping.yaml
│   ├── multi_brains_setting.yaml
│   └── README.MD
├── language
│   ├── en.yaml
│   └── zh.yaml
├── launch
│   └── llm_agent_control.launch.py
├── multi_brains
│   ├── action_service.py
│   ├── asr_detect.py
│   ├── __init__.py
│   ├── model_service.py
│   └── utils
├── package.xml
├── resource
│   └── multi_brains
├── setup.cfg
├── setup.py
├── system_vioce
│   ├── en
│   ├── notify.mp3
│   ├── test_en.wav
│   ├── test_zh.wav
│   └── zh
└── test

• config

Configuration folder, used to store configuration file templates.

• multi_brains source code folder
• asr_detect.py

Speech recognition program file

• model_service.py

Model server program file, used to call various model interfaces to implement the model inference architecture

• action_service.py

Action server program file, used to receive the action list requested by the model server and control the robot's movement

• launch folder

Folder for storing ROS2 node launch files

• language package

Stores log files in different languages

• system_voice

System audio files

3. Speech Recognition Function

Source code path:

xxxxxxxxxx
~/yahboomcar_ros2_ws/yahboomcar_ws/src/multi_brains/multi_brains/asr_detect.py

3.1 Dynamic Recording with Voice Activity Detection

1. Continuously read audio frames and perform voice activity detection.
2. If voice activity is detected, add the audio frames to the buffer; if continuous silence exceeding a threshold (90 frames, approximately 1.5 seconds) is detected, end the recording.
3. After recording ends, remove the trailing silent portion and save the valid speech as a WAV file.
4. If no valid speech is detected, no file will be saved.

​x
    def listen_for_speech(self):
        ''' Dynamic recording with VAD'''
        self.record_flag = True
        PRE_SPEECH_FRAMES  = 5         #150ms voice start compensation
        PRINT_EVERY_N_FRAMES = 5 
​
        recording_active = False
        silence_counter  = 0
        print_counter    = 0
​
        audio_buffer = []
        pre_speech_buffer = deque(maxlen=PRE_SPEECH_FRAMES)
        stream = self.pyaudio.open(
            format=pyaudio.paInt16,
            channels=1,
            rate=self.sample_rate,
            input=True,
            frames_per_buffer=self.frame_bytes,
        )
        self.play_audio(self.notify_audio)  
        try:
            while not self.stop_event.is_set():
                frame = stream.read( self.frame_bytes, exception_on_overflow=False)
                is_speech = self.vad.is_speech(frame, self.sample_rate)
                print_counter += 1
                if print_counter >= PRINT_EVERY_N_FRAMES:
                    print("1-1-1" if is_speech else "-----")
                    print_counter = 0
                if not recording_active:
                    # ---- IDLE ----
                    pre_speech_buffer.append(frame)
                    if is_speech:
                        #  RECORDING
                        recording_active = True
                        audio_buffer.extend(pre_speech_buffer)
                        pre_speech_buffer.clear()
                        audio_buffer.append(frame)
                        silence_counter = 0
                else:
                    # ---- RECORDING ----
                    audio_buffer.append(frame)
                    if is_speech:
                        silence_counter = max(0, silence_counter - 1)
                    else:
                        silence_counter += 1
                        if silence_counter >= self.MAX_SILENCE_FRAMES:
                            break
        finally:
            stream.stop_stream()
            stream.close()
            self.record_flag = False
​
        if recording_active and audio_buffer:
            with wave.open(self.user_speech_dir, "wb") as wf:
                wf.setnchannels(1)
                wf.setsampwidth(self.pyaudio.get_sample_size(pyaudio.paInt16) )
                wf.setframerate(self.sample_rate)
                wf.writeframes(b"".join(audio_buffer))
            return True
        return False

3.3 ASR Speech Recognition

• The recognize method of the speech engine is called for speech recognition. The speech engine is provided by instantiated objects of the classes Tongyi_ASR, SenseVoiceSmall_ASR, and XUNFEI_ASR.

xxxxxxxxxx
    def kws_handler(self) -> None:
        ''' Wake-up handling function'''
        if self.stop_event.is_set():
            self.logger.info("Wake-up processing thread interrupted  no handle user speech.!!!!!")
            return
        if self.listen_for_speech():
            asr_result=self.asr_engine.recognize(self.user_speech_dir)  #  Perform ASR conversion
            if not asr_result[0]:
                self.logger.error(Fore.RED+f"Speech recognition failed because the audio segment is empty or the speech model is unavailable.ASR_OUT:{asr_result[1]}"+Fore.RESET)
            else:
                if len(asr_result[1]) > self.asr_threashold:
                    self.logger.info(Fore.GREEN+"ASR Result: "+asr_result[1]+Fore.RESET)
                    self.asr_result_queue.put([asr_result[1],'text_request',False])#  Put the ASR result into the queue
                else:
                    self.logger.info(Fore.YELLOW+"The voice recognition result is too short. Could it be that the user woke it up by mistake "+asr_result[1]+Fore.RESET)

4. Model Service Functionality

Source code path:

xxxxxxxxxx
~/yahboomcar_ros2_ws/yahboomcar_ws/src/multi_brains/multi_brains/model_service.py

Core function explanation:

• Checks the llm_handler_queue queue in a separate thread for requests to access the model.
• Once an element is pushed into the queue, it calls the chat function of the Dify access interface with different parameters depending on the request type.
• After receiving the AI agent's response from the Dify platform, it parses the corresponding content for speech playback and sends the action list to the action server for execution.

xxxxxxxxxx
    def handle_llm_thread(self)->None:
        '''
         Handle model request
        '''
        while True:
            if not self.llm_handler_queue.empty():#The queue is not empty, processing model requests.
                if  not self.text_chat_mode and self.asr_detect.record_flag : continue
                request_query = self.llm_handler_queue.get()
                if self.debug_mode: self.get_logger().info(f"Processing LLM request: {request_query}")
​
                if request_query[1]=='text_request':
                    '''text request'''
                    result=self.dify_llmclient.chat(request_query[0],robot_feedback=request_query[2])
                elif request_query[1]=='image_request':
                    '''vision + text request'''
                    result=self.dify_llmclient.chat(request_query[0],image_path=self.image_cache_path,robot_feedback=request_query[2])
​
                if result[0]:
                    if  not self.text_chat_mode and self.asr_detect.record_flag : continue
                    split_result=self.extract_actions(result[1])
                    if split_result is None: continue
                    action_list,llm_response,decision_plan=self.extract_actions(result[1])
                    if decision_plan is not None: 
                        self.get_logger().info(Fore.YELLOW+self.syslog.get_text("system_log_3",decision_plan=decision_plan)+Fore.RESET)
                    self.get_logger().info(Fore.YELLOW+f'"action": {action_list},"response": {llm_response}'+Fore.RESET) 
​
                    if  not self.text_chat_mode:#Voice reply
                        if self.tts_engine.synthesize(llm_response,self.tts_out_path) :
                            self.play_audio(self.tts_out_path)
                        else:
                            self.get_logger().error(Fore.RED+"Speech synthesis failed. Check whether the TTS model is available"+Fore.RESET)
                    else:#Text Reply
                        if decision_plan is not None: 
                            self.text_pub.publish(String(data=self.syslog.get_text("system_log_3",decision_plan=decision_plan)))
                        self.text_pub.publish(String(data=f'"action": {action_list}, "response": {llm_response}'))
​
                    if action_list!=[]: self.send_action_service(action_list, llm_response)  
                else:
                    self.get_logger().error(Fore.RED+f"The model request failed. Check whether the dify or AI model is normal.\
                                            Error Log:{result[1]}"+Fore.RESET)
            else:  
                time.sleep(1.0)#Sleep for 1 second when there are no requests.

5. Action Server Functionality

• Contains the implementation of all basic action functions that the robot can perform. It receives action list requests and executes the corresponding actions. It also has the functionality to interrupt and resume execution after being reactivated.

Core callback function execute_callback explanation:

• Accepts a string representing the list of actions.

xxxxxxxxxx
    def execute_callback(self, goal_handle):
        """action execution callback function"""
        actions = goal_handle.request.actions
        feedback_result = None
        if self.debug_mode: self.get_logger().info(self.actionlog.get_text("debug_log_1",actions=actions))
        self.action_runing = True
        for action in actions:
            if self.interrupt_event.is_set():
                break
            match = re.match(r"(\w+)\((.*)\)", action)
            action_name, args_str = match.groups()
            args = [arg.strip() for arg in args_str.split(",")] if args_str else []
            if not hasattr(self, action_name):
                self.get_logger().error(Fore.RED+f"action_service: {action} is invalid action, skip execution"+Fore.RESET)
            else:
                method = getattr(self, action_name)
                feedback_result = method(*args)
​
        if not self.interrupt_event.is_set():#Provide feedback to the Dify agent on the action execution results
            msg=LlmRequest()
            if feedback_result==False:
               #Action failed
                msg.llm_request=self.actionlog.get_text("action_feedback_2",action_name=actions)
                msg.robot_feedback=True
                self.llm_request_pub.publish(msg)
            elif feedback_result==True:
                #Action executed successfully
                msg.llm_request=self.actionlog.get_text("action_feedback_1",action_name=actions)
                msg.robot_feedback=True
                self.llm_request_pub.publish(msg)
            elif feedback_result==None:
              #No operation, no feedback
                if self.debug_mode: self.get_logger().info(self.actionlog.get_text("system_log_1"))
​
            if self.debug_mode: self.get_logger().info(msg.llm_request)
​
        if self.debug_mode: self.get_logger().info(self.actionlog.get_text("system_log_2"))
        self.action_runing = False
        self.interrupt_event.clear()
        goal_handle.succeed()
        result = Rot.Result()
        result.success = True
        return result

6. Interruption Function

The robot supports interruptions at any stage, which can be specifically divided into recording stage interruptions, dialogue stage interruptions, and action stage interruptions. The principles of interruption at each stage are introduced below.

6.1 Recording Stage Interruption

If you realize you've made a mistake while recording, or are dissatisfied with the recorded content and need to re-record, you can interrupt the previous recording and start speaking and recording again by re-activating the robot during the recording process.

• The logic is implemented in the asr_detect_run method in the asr.py file:
• Each time the robot is activated, if there is already a thread running for wake-up recording, it is interrupted via the thread event stop_event, and waits for it to end;
• After clearing the stop event flag, a new recording thread is started.

xxxxxxxxxx
    def asr_detect_run(self):
        while True:
            #  Process only the most recent wake-up request to prevent duplicates
            if self.wakeup_event.wait(timeout=0.1):
                self.wakeup_event.clear()
                self.extern_wakeup.set()
                self.publisher.wakeup_pub.publish(Bool(data=True))
                
                self.logger.info("I'm here🌟")
                self.wake_up_voice() #  Respond to the user
                if self.current_thread and self.current_thread.is_alive():   #  Interrupt the previous wake-up handling thread
                    self.stop_event.set()
                    self.current_thread.join()  # Wait for the current thread to finish
                    self.stop_event.clear()  #  Clear the event
                self.current_thread = threading.Thread(target=self.kws_handler)
                self.current_thread.daemon = True
                self.current_thread.start()
            time.sleep(0.5)

6.2 Interrupting the Dialogue Phase

If you are dissatisfied with the robot's response during its speech or don't want the robot to continue speaking, you can use the wake word to interrupt the robot's speech and start recording your voice. At this point, you can give the robot a new command (still within the current task cycle), or you can say "End current task" to directly end the current task and start a new task cycle.

• The logic is implemented in the wakeup_callback and play_audio methods of the CustomActionServer class in the action_service.py file:
• wakeup_callback is the callback function for wake-up processing. In the asr.py program, each time a wake-up signal is detected, it publishes a wake-up signal via topic communication. wakeup_callback subscribes to and processes this signal.
• Each time a wake-up occurs, it checks whether pygame.mixer is currently playing audio. If so, it notifies the playback thread to stop playback via the thread event self.stop_event.
• If a previous action is detected to be running after the wake-up, the self.interrupt_flag flag is set, which is used for subsequent action interruption.

xxxxxxxxxx
def wakeup_callback(self, msg):
    if msg.data:
        if pygame.mixer.music.get_busy():
            self.stop_event.set()
        if self.action_runing:
            self.interrupt_flag = True
            self.stop()
            self.pubSix_Arm(self.init_joints)

When play_audio is playing audio, it checks if self.asr_detect.extern_wakeup is set. If it detects that it is set, it immediately stops the currently playing audio.

xxxxxxxxxx
    def play_audio(self,file_path: str) -> None:
        '''Play audio'''
        self.asr_detect.extern_wakeup.clear()
        with self.pygame_lock:
            pygame.mixer.init()
            pygame.mixer.music.load(file_path)
            pygame.mixer.music.play()
            while pygame.mixer.music.get_busy():
                if self.asr_detect.extern_wakeup.is_set():
                    pygame.mixer.music.stop()
                    self.asr_detect.extern_wakeup.clear()
                    break
                pygame.time.Clock().tick(10)
            pygame.mixer.quit()

6.3 Action Phase Interruption

If the robot is interrupted during the execution of an action, it will stop the current action and return to its initial posture. This can be divided into two types: normal action interruption and action interruption with subprocesses.

6.3.1 Normal Action Interruption

• The robot's chassis movement and robotic arm movement are controlled by publishing velocity topics and robotic arm joint angle topics.
• The _execute_action chassis control function continuously checks the self.interrupt_event interruption flag. If it is set, the chassis movement is immediately stopped.
• Similarly, the pubSix_Arm robotic arm control function checks the self.interrupt_event interruption flag; it will only publish the robotic arm joint angle topics normally if the flag is not set.

xxxxxxxxxx
def _execute_action(self, twist, num=1, durationtime=3.0):
    for _ in range(num):
        start_time = time.time()
        while (time.time() - start_time) < durationtime:
            if self.interrupt_flag:
                self.stop()
                return
            self.publisher.publish(twist)
            time.sleep(0.1)
            
def pubSix_Arm(self, joints, id=6, angle=180.0, runtime=2000):
    arm_joint = ArmJoints()
    arm_joint.joint1 = joints[0]
    arm_joint.joint2 = joints[1]
    arm_joint.joint3 = joints[2]
    arm_joint.joint4 = joints[3]
    arm_joint.joint5 = joints[4]
    arm_joint.joint6 = joints[5]
    arm_joint.time = runtime
    if not self.interrupt_flag:
        self.TargetAngle_pub.publish(arm_joint)

6.3.2 Interruption of Actions with Subprocesses

For example, actions such as robotic arm gripping and sorting machine codes require starting external programs in subprocesses. Here, we'll use the robotic arm gripping action function grasp_obj as an example:

When the robotic arm gripping is not complete, it will continuously wait in a while not self.grasp_obj_future.done(): loop. During this process, if the self.interrupt_event flag is detected as set, the corresponding __reset_grasp_obj() function will be called first, recursively ending the subprocess tree, and then the action will stop.

xxxxxxxxxx
    def grasp_obj(self, x1, y1, x2, y2) -> None:
        """grasp_obj: Grasping objects x1,y1,x2,y2: Object outer border coordinates """
        def __reset_grasp_obj():
            kill_process_tree(self.grasp_obj_process_1.pid)
            kill_process_tree(self.grasp_obj_process_2.pid)
            kill_process_tree(self.grasp_obj_process_3.pid)
            self.grasp_obj_future = Future() 
            
        cmd_1=['ros2', 'run', 'largemodel_arm', 'grasp_desktop']
        cmd_2=['ros2', 'run', 'largemodel_arm', 'KCF_follow']
        cmd_3=['ros2', 'run', 'M3Pro_KCF', 'ALM_KCF_Tracker_Node']
​
        self.grasp_obj_process_1=subprocess.Popen(cmd_1)
        time.sleep(5.0) #Waiting for grasp_desktop to finish starting up
        self.grasp_obj_process_2=subprocess.Popen(cmd_2)
        self.grasp_obj_process_3=subprocess.Popen(cmd_3)
        x1 = int(x1)
        y1 = int(y1)
        x2 = int(x2)
        y2 = int(y2)
        while not self.object_position_pub.get_subscription_count():
            time.sleep(0.5)
        self.object_position_pub.publish(Int16MultiArray(data=[x1, y1, x2, y2]))
​
        while not self.grasp_obj_future.done():
            if self.interrupt_event.is_set():
                __reset_grasp_obj()
                self.pubSix_Arm(self.init_joints)
                return None
            time.sleep(0.1)
​
        result = self.grasp_obj_future.result()
        if not self.interrupt_event.is_set():
            if result.data == "grasp_obj_done":
                res = True
            else:
                res = False
​
        __reset_grasp_obj()
        if self.interrupt_event.is_set():
            time.sleep(0.5)
            self.pubSix_Arm(self.init_joints)  # Robotic arm retracted
        return res